Calculating Strain on a Wireless Telecommunication Network

962
4
06-04-2017 02:57 AM
SimonJackson
Occasional Contributor III

I have four datasets - point layers with different measurement variables, covering the same extent.

I think a raster based approach makes sense (although will also be investigating a vector binning approach later on) so I have calculated interpolations upon these different datasets.

The results are 4 continuous surfaces (floating point).
They share the same area, but some have holes in the surfaces (which is expected).

  • Quality of Experience (~0-20)
  • Radio Metrics (~0-150)
  • User density (~0-60)

High Strain on the network = High Quality of Experience + High Radio Metrics
Low Strain on the network = Low Quality of Experience + Low Radio Metrics + High User Density

I was initially thinking this would be a good candidate for a weighted overlay, but then realised that I would have to reclassify the input datasets into discrete bands.  I want to avoid this and maintain the continuous values in the datasets, as I am hoping to output a strain layer that is also continuous (Strain will be a measure anywhere between 0-100).

I am now thinking I could potentially achieve what i want using standard raster calculator expressions within a model, but feel like I am missing something and there might be another set of overlay tools I should be looking at?

I was wondering if I could call upon the brains of this Analysis group to perhaps brainstorm some approaches for the above scenario. 

This is just an initial investigation project.  Longer term I would actually like to investigate how the new Raster Analytics Server might be able to play a part in this workflow as the underlying datasets are changing frequently. 

0 Kudos
4 Replies
DanPatterson_Retired
MVP Emeritus

Quality of experience - nominal or at best ordinal data

Radio metrics -             same as above or are the data quantifiable on an interval/ratio scale

User Density -              ratio, I presume if it is users/unit area or equivalent

High Strain on the network = High Quality of Experience + High Radio Metrics

    Can't add nominal scales or ordinal scales together or with any other measurement scale

Low Strain on the network = Low Quality of Experience + Low Radio Metrics + High User Density

   Ditto

You are back to working with non-parametric measures of association.  Combine (as previously discussed) may with a bit of 'fuzzification' if your boundaries are not discrete.

Don't upscale data, use an analytical approach suitable and supported by your data measurement scale.  There is nothing wrong with this and is sure better using the wrong methodology.

As for 'weighted anything' ... is it supported by fact, observation or conjecture?  I would test for it first.

Just some thoughts.

SimonJackson
Occasional Contributor III

Hey Dan - appreciate your thoughts.

All three datasets are I believe interval data and not nominal or ordinal.  Let me edit this answer in a bit with some summary stats and screenshots to give you a better feel for the datasets.  This might change your answer?

As for 'weighted anything' ... is it supported by fact, observation or conjecture

We will be doing some statistical testing in order to come up with the weightings.

0 Kudos
DanPatterson_Retired
MVP Emeritus

Simon... some more

Quality of experience might at best be ordinal...ie ranked.  It isn't something that you can measure directly except via surrogate variables or opinion and an 'experience = 10' can vary widely from person to person, whereas 10 degrees centigrade is always the same... similarly in order to be interval, the spacing between the values must be equal... ordinal data has not expectation or requirement of this.

0 Kudos
SimonJackson
Occasional Contributor III

Can understand why you would assume QoE would be ordinal as sounds like something you would survey from people, but it is actually a measure calculated from download throughput.

I do appreciate this is still very early stages in the data exploration, and definitely treading carefully.  Still toying between a raster vs vector based analysis.  The below animated gif might give you a better indication of the data I am working with.  Still needs a lot of work, some outliers to remove, etc, but would still be very interested to hear additional thoughts you might have.

0 Kudos